@agent-runtime-map/mcp

A Model Context Protocol server that lets any agent read a repository as an evidence-backed map instead of file by file.

Register it

{
  "mcpServers": {
    "agent-runtime-map": {
      "command": "npx",
      "args": ["-y", "@agent-runtime-map/mcp"]
    }
  }
}

Until the package is on npm, point at a local build instead:

{
  "mcpServers": {
    "agent-runtime-map": {
      "command": "node",
      "args": ["/absolute/path/to/agent-runtime-map/packages/mcp/dist/index.js"]
    }
  }
}

Tools

| Tool | Answers | | --- | --- | | analyze_project | What is this repository, and what capabilities does it have? | | list_features | Which feature circuits exist, and which are broken? | | describe_feature | What are the steps of one feature, and how confident is each? | | get_evidence | What was this step read from, and what does the project claim about it? |

Each answers one question and names the tool that answers the next, because an agent pays for every token it reads. A compiled map of a real repository is hundreds of kilobytes, and no tool returns it whole.

What the answers keep

Every step carries its source location and a confidence, and a step matched to a documented capability reports that separately from its code confidence. A map without those is an architecture poster, and an agent repeating it would be stating as fact something nobody can check.

A route is a statically inferred path, and every answer that shows one says so. It is not a recording of anything that ran.

What it will not do

• It does not execute the project. Analysis reads source. The Python adapter shells out to ast.parse, which builds a tree and imports nothing.
• It writes nothing unless analyze_project is called with write: true.
• It is not a file reader. get_evidence can return source lines, but only for a path the graph already points at, only inside the analyzed project, and only a bounded number of lines.
• It does not call a model. Optional LLM enrichment exists in the CLI and is off here.